Start with the research, including its limits.
The EBU report on news integrity in AI assistants examines problems in generated news answers, including sourcing and accuracy. Read its methods and examples before applying the findings to a different model, question, or point in time.
The Reuters Institute Digital News Report 2025 provides a different perspective: how people use and judge news. Audience concern and model error rates measure different things; one cannot stand in for the other.
Read the citation before using the answer.
Open the original record. Does it exist? Does the passage support the sentence? Does the date match the claim? A source can be real and relevant without establishing the conclusion attributed to it.
A useful distinction
An answer may quote a source correctly but apply it to the wrong population, period, or definition. Check the scope as well as the wording.
Make the next check specific.
- Write down the exact claim, including its date and qualifiers.
- Find the primary source and the passage that addresses it.
- Look for a later update or an independent source that changes the conclusion.
- Keep unresolved details in the final wording.
Keep a reviewable record.
Save the source link, relevant passage, and any limitation that affected your decision. The NIST AI Risk Management Framework offers broader guidance on managing AI risks; it does not certify an individual answer.
This checklist is Verity’s editorial guidance. It is not an evaluation of a particular deployed model or a measured claim about Verity’s accuracy.